TypeSafe AI's Jev 1.14 Brings Vision and 256K Context to Typed Decisions
TypeSafe AI ships Jev 1.14 with image inputs, a bounding box primitive, and 4x longer context, all behind opt-in danger zone headers.
- TypeSafe AI released Jev 1.14.0 with across-the-board improvements, biggest gains in multilingual support.
- Image inputs now work anywhere text does, via composable placeholders in state, instructions, or criteria.
- New Box primitive returns bounding box coordinates as fractions of image dimensions.
- Context length cap quadrupled when the long-context danger zone header is enabled.
- Experimental features gated behind opt-in
TypeSafe-Danger-ZoneHTTP headers, not production certified. - Pricing unchanged at $42 per billion input tokens; output tokens remain free.
TypeSafe AI has released Jev 1.14.0 with three experimental capabilities for its System One decision model: images in any text field, normalized bounding-box answers, and context limits four times larger than the defaults. TypeSafe places all three in its opt-in “danger zone” program because they have not met the company’s production reliability threshold.
Jev exposes one API endpoint, POST /v1/systemone, for asking typed questions about shared state. Callers request booleans, choices, scores, and other schema-constrained values that automation code can consume directly. Version 1.14 preserves that structure while adding visual inputs, spatial outputs, and more input capacity.
Three opt-in switches
- Images: Place image references in state, instructions, or criteria. Enable them with
TypeSafe-Danger-Zone-Images: true. Image support requires the TypeSafe-Danger-Zone-Images header and is available for Jev 1.13.0 and 1.14.0, though Box questions specifically require Jev 1.14. - Box: Request the normalized coordinates of an object in an image. The experimental Box primitive requires Jev 1.14.
- Long context: Multiply the standard context limits by four with
TypeSafe-Danger-Zone-Long-Context: true.
Images become reusable assets
System One resolves image placeholders inside state, question instructions, and criteria. Each image is declared once in a top-level assets map keyed by its placeholder, allowing several questions to reuse the same asset. A request can define Choice options with example images or place a short frame sequence in state for motion analysis.
The underlying Jev 1.13 model already supported visual inputs, so TypeSafe is exposing the feature on both 1.13 and 1.14. The company warns that image tasks remain less reliable than text tasks and recommends splitting complex visual judgments into focused questions.
| Image constraint | Limit |
|---|---|
| Formats | PNG and JPEG still images; animated PNGs are rejected |
| File size | 16 MiB per image |
| Dimensions | 256 to 1,568 pixels on each side |
| Request body | 32 MiB total, including JSON, base64 data, and multipart framing |
| Billing | Images count as input tokens at the same per-token rate as text |
Box always returns coordinates
Box adds a spatial output type to System One. A request specifies type: "box" and an instruction such as the coffee mug in photo. Jev returns top, bottom, left, and right as values from 0 to 1, along with a confidence score.
The coordinates are fractions of the source image dimensions. Callers can recover pixel positions by multiplying left and right by the image width, and top and bottom by its height.
Box has no absent-object state, so it emits four coordinates even when the requested object is missing. TypeSafe recommends pairing it with a Noul boolean question that checks whether the target exists, then consuming the Box result only when that check succeeds.
TypeSafe’s documentation reports a cutoff-detection evaluation using 36 cropped paintings, each with a known portion of the target outside the crop:
| Method | Cuts detected |
|---|---|
| Four Noul questions, one for each edge | 30 of 36 |
| One Noul question asking whether the whole object is inside the crop | 26 of 36 |
| Draw the returned Box as an outline on the full image | 20 of 36 |
The figures cover one small task, but they provide a concrete reason to decompose edge-sensitive checks instead of relying on a single broad question.
The window expands to 256k
Jev’s default budget remains 64,000 tokens across a request, with a separate 32,000-token limit for the shared state plus the longest individual question. The long-context header multiplies both limits by four.
| Budget | Default | Long context |
|---|---|---|
| Complete request | 64,000 tokens | 256,000 tokens |
| State plus longest question | 32,000 tokens | 128,000 tokens |
TypeSafe says Jev can technically accept the larger input, but reliability declines as context grows. The lower default encourages smaller state objects and narrower questions; workloads that enable the expanded window need task-specific evaluations at their expected input lengths.
Preview moves first
Jev 1.14.0 also includes general model updates, with the largest reported gains in multilingual tasks. The jev-preview alias now points to 1.14.0, and TypeSafe plans to move jev-latest one week after the release. Clients that require repeatable behavior can pin the exact model version.
TypeSafe has published internal benchmark results against open-weight comparison models while advising customers to run evaluations on their own workloads. Pricing remains $42 per billion input tokens, equivalent to $0.042 per million, with output tokens free.
The release broadens the typed decision API into visual input and spatial output while retaining a single request model. Ticket classifiers can include screenshots, frame-triage systems can return presence and location, and document workflows can produce normalized crop coordinates. Production use still requires careful testing of image accuracy, missing-object handling, and long-context degradation.
Comments
